[Synth ] feat : Fold Majority Graph #9033

markram1729 · 2025-09-27T07:19:33Z

Resolves Partial #8918
// Folder for
maj_inv(x, !x, y) -> y
maj_inv(x, x, y) -> x
maj_inv(x, 0, 1) -> x

uenoku · 2025-09-28T08:54:38Z

lib/Dialect/Synth/SynthOps.cpp

+        // If i and j are constant.
+        if (auto c1 = getConstant(i)) {
+          if (auto c2 = getConstant(j)) {
+            // If both constants are equal, we can fold.
+            if (*c1 == *c2) {
+              // auto value = cast<IntegerAttr>(getInputs()[i]).getValue();
+              //  return IntegerAttr::get(IntegerType::get(getContext(),
+              //  value.getBitWidth()),value);
+              return IntegerAttr::get(getType(), *c1);
+            }
+            // If constants are complementary, we can fold.
+            if (*c1 == ~*c2)
+              return getOperand(k);
+          }
+        }


This block would be dead as two constants are already handled.

yes handled !

uenoku · 2025-09-28T08:54:44Z

lib/Dialect/Synth/SynthOps.cpp

  auto width = inputs[0].getBitWidth();
  APInt result(width, 0);

+  // is every input of same width ?


uenoku · 2025-09-28T08:56:10Z

lib/Dialect/Synth/SynthOps.cpp

+    // x 1 1
+    if (inputValues.size() == 2) {
+      if (inputValues[0] != inputValues[1])
+        return {};


I think it's pessimistic. If two constants are same you can return constants, if they are complement you can return third variable.

true , resolved !

uenoku · 2025-09-28T08:57:04Z

lib/Dialect/Synth/SynthOps.cpp

+          if (isInverted(i) != isInverted(j)) {
+            return getOperand(k);
+          }
+          return getOperand(i);


This is a my bug but it's necessary to check if operand i is inverted. Could you fix it as well?

Hi @uenoku ,
if operand(i) is inverted , should i return {} ?

I can't return with inverted operand(i) , let canonicalise handle it !

uenoku

Thanks!

uenoku · 2025-09-28T08:59:27Z

lib/Dialect/Synth/SynthOps.cpp

+          }
+        }
+      }
+    }


you need to return {} here as evaluate expects same operands. Also evaluate expects uninverted values so the current code is not correct.

yes , it is handled now

Thank you for pointing out

markram1729 · 2025-10-01T09:00:44Z

Hi @uenoku , Can you please check if this resolves the folding issue ?

uenoku · 2025-10-02T03:17:47Z

lib/Dialect/Synth/SynthOps.cpp

  assert(inputs.size() == getNumOperands() &&
         "Number of inputs must match number of operands");

  if (inputs.size() == 3) {


Could you keep the previous implementation? evaluate is meant to evaluate the op based on input values so it's necessary to look at inversion.

uenoku

Sorry for delayed review!

uenoku · 2025-10-02T03:20:44Z

lib/Dialect/Synth/SynthOps.cpp

+            IntegerType::get(getContext(), inputValues[0].getBitWidth()),
+            inputValues[0]);
+      else
+        return getOperand(index); // fix if it is a Inverted


Could you check the inversion?

uenoku · 2025-10-02T03:28:51Z

lib/Dialect/Synth/SynthOps.cpp

-  // Pattern match following cases:
-  // maj_inv(x, x, y) -> x
-  // maj_inv(x, y, not y) -> x


Could you keep the canonicalization since folder is essentially a bit weaker?

uenoku · 2025-10-02T03:43:28Z

lib/Dialect/Synth/SynthOps.cpp

  SmallVector<APInt, 3> inputValues;
+  size_t i = 0, index;


Could you do slightly change the code structure to store non-constant index and check based on the number constants?

Suggested change

SmallVector<APInt, 3> inputValues;

size_t i = 0, index;

SmallVector<APInt, 3> inputValues;

SmallVector<unsigned, 3> nonConstantIndex;

for (auto [i, input] : llvm::enumerate(adaptor.getInputs())) {

auto attr = llvm::dyn_cast_or_null<IntegerAttr>(input);

if (attr)

inputValues.push_back(attr);

else

nonConstantIndex.push_back(i);

}

if (inputValues.size() == getNumINputs())

return IntegerAttr::get(getType(), evaluate(inputValues));

if (getNumOperands() != 3)

return {};

if (nonConstantIndex.size() == 1) {

// Get two constants.

auto idx = nonConstantIndex.front();

auto constIndex1 = (idx+1)%3;

auto constIndex2 = (idx+2)%3;

auto c1 = adaptor.getInputs()[constIndex1];

auto c2 = ...

auto ...;

}

markram1729 · 2025-10-02T06:54:16Z

Hi @uenoku ,
I have pushed the new changes , I will optimise it more

Can you suggest how can i handle ~x in fold

any way i can return ~x as i can't negate a operand
will canonicalise handle it ? I am returning as return {}

uenoku · 2025-10-03T00:42:40Z

Really thank you for working hard on this 🙇🙇!

Can you suggest how can i handle ~x in fold

Unfortunately we cannot handle it since it's not allowed to create a new operation: https://mlir.llvm.org/docs/Canonicalization/#canonicalizing-with-the-fold-method
So it's necessary to just give up the folding for such cases.

markram1729 · 2025-10-03T06:19:46Z

Hi @uenoku , can i handle in fold ?

 x x 1 -> x
 x ~x 1 -> 1
 ~x ~x 1 -> ~x

I have commented in the code ,as canoicalisation ishandling it

uenoku · 2025-10-05T08:49:22Z

The first two can be implemented in a straightforward way.

For the 3rd one, i realized maybe MLIR folder allows in-place mutation for the op itself, so maybe you can do something like

            (*this)->setOperands({getOperand(k)});
            (*this).setInverted({true});
            return getResult();

markram1729 force-pushed the exaSynthMaj branch from aed800d to d5dcd5b Compare September 27, 2025 07:27

uenoku reviewed Sep 28, 2025

View reviewed changes

markram1729 force-pushed the exaSynthMaj branch 5 times, most recently from 7c95602 to 272d8e2 Compare September 29, 2025 07:24

markram1729 marked this pull request as ready for review September 29, 2025 07:26

markram1729 requested a review from uenoku September 29, 2025 15:02

uenoku reviewed Oct 2, 2025

View reviewed changes

markram1729 marked this pull request as draft October 2, 2025 06:38

markram1729 requested a review from uenoku October 3, 2025 06:32

markram1729 added 8 commits October 3, 2025 12:06

feat : Fold Majority Graph

f0db3f5

fix : foldop

392392b

opt : foldop

84ba248

fix : Inverted Op

4540f7e

chore : cleanup

ca87df7

refactor : FoldOp

f7078a6

fix : names

1574dd1

fix error : foldOp

ff77ae4

markram1729 force-pushed the exaSynthMaj branch from 74a6f8e to ff77ae4 Compare October 3, 2025 06:37

fix : ~x const fold

dd5f31b

-  SmallVector<APInt, 3> inputValues;
-  size_t i = 0, index;
+  SmallVector<APInt, 3> inputValues;
+  SmallVector<unsigned, 3> nonConstantIndex;
+ for (auto [i, input] : llvm::enumerate(adaptor.getInputs())) {
+    auto attr = llvm::dyn_cast_or_null<IntegerAttr>(input);
+    if (attr)
+        inputValues.push_back(attr);
+     else
+       nonConstantIndex.push_back(i);
+  }
+  if (inputValues.size() == getNumINputs())
+    return IntegerAttr::get(getType(), evaluate(inputValues));
+  if (getNumOperands() != 3)
+    return {};
+ if (nonConstantIndex.size() == 1) {
+    // Get two constants.
+    auto idx = nonConstantIndex.front();
+    auto constIndex1 = (idx+1)%3;
+    auto constIndex2 = (idx+2)%3;
+    auto c1 = adaptor.getInputs()[constIndex1];
+    auto c2 = ...
+    auto ...;
+ }

[Synth ] feat : Fold Majority Graph #9033

Are you sure you want to change the base?

[Synth ] feat : Fold Majority Graph #9033

Conversation

markram1729 commented Sep 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

uenoku Sep 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

markram1729 Sep 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

uenoku left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

markram1729 commented Oct 1, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

uenoku left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

markram1729 commented Oct 2, 2025

Uh oh!

uenoku commented Oct 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

markram1729 commented Oct 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

uenoku commented Oct 5, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

markram1729 commented Sep 27, 2025 •

edited

Loading

uenoku Sep 28, 2025 •

edited

Loading

markram1729 Sep 29, 2025 •

edited

Loading

uenoku commented Oct 3, 2025 •

edited

Loading

markram1729 commented Oct 3, 2025 •

edited

Loading